AXHUB LECTURE · INTERMEDIATE L5

Before you hand off to an agent
the 5 checks — hands-on

Time about 30 minutesWhat you need one task you want to automate (or already run)For anyone who finished L1–L4, and users of automation tools

Building an agent now takes half a day. The problem is what comes after — there's a real war story going around where one line, "fix the code for me," made production data vanish in 9 seconds. Before you build, fill in five boxes. The output of this lecture isn't code but a single document.

1Permissions — how far will you let it reach

An agent does what you ask "too diligently."

So permissions are set by a person, not the task. Step by step: read first, write after review, delete never.

Practice 1 — fill in the permission boxes
My automation: ____
May read: ____ (e.g. the inquiry inbox)
May write: ____ (e.g. the draft-reply folder)
Never allowed: ____ (e.g. sending, deleting, paying)

Check: is the "never allowed" box not left empty?

2Recovery — can it be undone

Some things can't be taken back once they're out. Sending, paying, deleting, posting publicly.

For those, let it go "only up to the last step." A person presses the final button.

Even Meta's executive in charge of AI safety went through it — an agent set to "confirm before executing" deleted 200 messages from the inbox and ignored the stop command too (2026, disclosed by the person themselves). Instructions are not a guardrail. Permissions are the guardrail.

Practice 2 — make a stop-before rule
Irreversible step: ____
How far the agent goes: ____ (e.g. up to the draft folder)
Backup/retention rule: ____ (e.g. move to an archive instead of deleting)

Check: imagining the worst malfunction, is there a recovery path?

3Exceptions — what has to come to a person

All-automatic and accidents pile up quietly. All-approval and there's no point in automating.

The answer is routing — automatic as usual, only the ambiguous ones to a person.

This structure outperformed approve-each by more than double (51 cases, +71% vs +30%). Accidents get filtered here too.

Practice 3 — exception criteria in 3 lines
May go through automatically: ____ (e.g. general inquiries)
Must always come to a person: ____ (e.g. refunds, complaints, any mention of an amount)
If ambiguous: no auto-send, to the hold queue

Check: does the "always to a person" list include items involving money, law, or emotion?

4Measurement — what will you measure "better" by

AI moves a little differently each time, even on the same instruction.

Fixing without a baseline turns improvement into a "gut feeling" — a shared lament among agent operators.

Executives say they save 8 hours a week while staff say 0 to 2 hours — the surveys say the same thing. Until you measure, both are just feelings.

Practice 4 — a one-line metric
My one-line metric: "of this week's ____, the share that went out without human edits"
This week's value: ____% → target in 4 weeks: ____%

One ratio is enough. Past two metrics, you stop measuring.

Check: is there a scheduled time next week, same weekday, to measure this number again?

5Records — leave it as a team asset

What made the agent run well isn't the model but the instructions and exception rules you refined.

Kept in a personal note, they vanish when a person leaves. Kept as a team document, they remain even if you switch tools.

77% of the gap comes not from technology but from operations like this.

Practice 5 — complete a one-page document
Title: ____ automation operations doc
1. Permissions (P1) 2. Recovery (P2) 3. Exceptions (P3) 4. Metric (P4)
5. Full instruction text + failure cases and how they were fixed

Paste in the four boxes you filled today and you're done. This one page is the output of this lecture.

Check: could a teammate (or future you) take over the operation from this document alone?

When it doesn't work

The "always to a person" list keeps getting longerThat's a normal sign. After a month, drop items that never came up, and add newly encountered accident types. The exception list is a living document.
I keep forgetting to measure the metricSet a measuring day and pin it as a recurring calendar event. And keep to just one metric — past two, you'll stop measuring.
I made the document but no one reads itDesignate it as the first document a new owner opens during handover. An operations doc not linked to onboarding is the same as no document.

From the start → L1. Start with one task you repeated today · Full contents

Sources: Stanford Enterprise AI Playbook (escalation +71% vs approve-each +30% · 77% of the hardest part is operations) · productivity-paradox surveys (Section, METR, PwC, Workday — the gap between perceived and measured) · Meta executive's inbox-deletion incident (own disclosure on X, multiple reports, 2026) · field war stories and operator reviews (social media anecdotes). Full sources are in the axhub.net case library. The practice items are templates to follow along.